Phased Approach | Are Open Source models finally ready for business use?

This week brought two significant advancements in generative AI, with Meta and Mistral releasing new open-source models that could rival industry leaders like GPT-4 and Claude 3.5 Sonnet.

On Tuesday, Meta's CEO Mark Zuckerberg introduced Llama 3.1, a model proclaimed to be on par with the most advanced proprietary systems from tech giants like OpenAI and Google.

The next day, Mistral, a rising star in the French AI scene, launched Mistral Large 2, touted for its superior multilingual capabilities and performance.

Business Impact

These releases could mark a shift in the AI landscape, offering businesses powerful AI tools at a fraction of the cost of proprietary models. Here’s a closer look at what this means for business applications.

What Was Released?

Llama 3.1 Meta's 405B model is the first openly available AI model that rivals top-tier alternatives. It excels in general knowledge, steerability, mathematics, tool usage, and multilingual translation. This model, along with upgraded versions of the 8B and 70B models, supports tasks like long-form text summarization and coding assistance. These models are available for download on llama.meta.com and Hugging Face, ready for immediate use on various partner platforms.

Mistral Large 2 Mistral's latest model features a 128k context window and supports numerous languages, including over 80 coding languages. Optimized for high throughput in long-context applications, this 123 billion parameter model is designed for single-node inference. It is available for research and non-commercial use under the Mistral Research License, with a commercial license available upon request.

Implications for AI in Business

In my work with companies, a lot of leaders have asked me about using Llama and Mistral but thus far adoption has been slow and disappointing. The reality is that the models just haven't been good enough. The lure of closed models has always been how easy it is to integrate an API into an existing process rather than maintaining a custom model. We are seeing that this is now getting easier and cheaper, so have we reached the tipping point.

Accelerating Innovation Open-source models like Llama 3.1 and Mistral Large 2 provide startups with high-quality, cost-effective AI tools. This can be a game-changer, allowing smaller companies to innovate without the heavy financial burden of proprietary AI models. With these tools, startups can focus on developing groundbreaking applications across various sectors, such as healthcare, finance, and retail.

Competitive Edge and Customisation These models offer the flexibility to build and customise solutions tailored to specific needs. Startups can modify and improve open-source models to develop unique features and capabilities that set them apart in the market.

Collaborative Development and Community Support The open-source nature of these models fosters a collaborative environment where developers and researchers can contribute to shared advancements. This community-driven approach accelerates the development of new features, providing startups with cutting-edge tools and insights to stay competitive.

Ethical and Security Considerations While democratising AI brings numerous benefits, it also raises ethical and security concerns. Startups must ensure their AI applications comply with industry standards and regulations, prioritising secure and ethical solutions to maintain trust and credibility.

Industry Impact The rise of open-source AI models compels industry leaders to continually innovate, emphasising unique features and seamless integration capabilities. This increased competition can create a dynamic AI landscape, benefiting businesses of all sizes and enabling startups to disrupt traditional industries with novel solutions.

Open Source in AI

What is Open Source in AI? In AI, "open source" ensures transparency, reproducibility, and unrestricted access to both the model and its components. Key elements include:

• Transparency: Full disclosure of data sources and model architecture.
• Reproducibility: Access to training data and procedures.
• Unrestricted Access: Free access to model weights and code.
• Community Involvement: Collaborative development and open governance.

Open Source AI Models vs. Proprietary Models Open-source models provide transparency and unrestricted access, while proprietary models often restrict access to data, weights, or code, usually for commercial or legal reasons.

Is Llama 3.1 Really Open Source? Llama 3.1 from Meta is not fully open source due to incomplete data transparency and potential usage restrictions. While Meta provides access to model weights, the lack of full disclosure of training data and licensing limitations means it doesn’t fully meet open-source criteria.

So are Open Source models finally ready for business use? Well only time will tell, when the new big release from Meta comes out we tend to see a lot of benchmark nerds making videos and writing articles about how this is a full paradigm shift. It never seems to actually happen though. The decision makers and implementors tend to shrug and move on with their projects using OpenAI and Anthropic. My sense is that this will eventually tip but there needs to be a compelling reason.

The reasons that are sited most commonly are Price and Privacy. Deploying Llama 3.1 or Mistral certainly isn't free, particularly when you factor in the overhead of having to set up and maintain environments. In terms of privacy, the closed models are aiming their premium offerings to business and so providing models that are not training on company data. My take is that there would have to be a specific industry reason to pick open source right now such as working with medical data in a proprietary environment or building consumer products with open models baked into the tech-stack.

I definitely will be watching this space to see if this really is the inflection point that Mark Zuckerberg is hoping for. He is spending the GDP of a small country to train these models so it is obviously a large bet for Meta. I don't think anyone long term would bet against Zuck. Closed models may still be the easier and frictionless option for most business applications though.